← ClaudeAtlas

research-code-skilllisted

Coding-standard enforcer for AI agents working inside a research (ML/DL) lab project. Its core job is twofold: (1) keep the project's fixed architecture intact across all coding work 鈥?never break or drift from the established structure; (2) enforce code-level conventions 鈥?naming, formatting, calling, and configuration 鈥?so every file is uniform. It codifies five authoritative systems as machine-executable Rule Cards (RC-*). It fuses thirteen authoritative references into four usage clusters so the agent never stacks them as separate checklists, plus a Karpathy-inspired behavioral discipline (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) layered above the mechanical cards: (1) Project Scaffold & Python Grammar (Lightning-Hydra-Template structure + Hydra configs + Google Python Style); (2) Model & Component Design (PyTorch Lightning Style Guide + timm named architectures + OpenMMLab Registry pattern); (3) Experiment Reproducibility (Hydra Config-First + FAIR data + SemVer/Git
SaltGardenia/research-code-skill · ★ 1 · Code & Development · score 70
Install: claude install-skill SaltGardenia/research-code-skill
# Research Code Skill > **For developers (human readers):** all documentation in this skill is written > in English, which is the source of truth. The agent always loads the **English** > originals. A skill that acts as the **coding-standard layer** of a research lab's ML/DL project. Rather than reviewing code after the fact, it lives in the agent's context and **constrains every code operation as it happens**. Its essence is twofold: 1. **Project architecture** 鈥?keep the lab's fixed project structure intact. As the agent writes or edits code, the architecture is the stable base: do not break it, do not invent a new layout, keep it clear and consistent. 2. **Code-level conventions** 鈥?within that architecture, every symbol and call follows uniform rules: naming, formatting, import/call style, and how configs compose. These two axes are enforced in both core scenarios (build from zero, tidy an existing repo). The agent consults this skill before writing or editing any project code, so the codebase stays coherent by construction. It distills thirteen authoritative references into **four fused usage clusters**, plus a coordination layer, into one shared convention set. Each cluster is one coherent concern the agent reaches for 鈥?the sources inside it compound, they do not stack: 1. **Project Scaffold & Python Grammar** (`references/scaffold_grammar.md`) 鈥?fuses Lightning-Hydra-Template (structure, `LHT-*`), Hydra configs (`HY-*`), and Google Python Style